[News Terms] Beyond Generative AI, 'Physical AI' Is Coming

Combination of Physical Laws and Data-Driven Learning
Jensen Huang Highlights Physical AI at CES 2025
Humanoid Robot Market Expected to Grow Rapidly

Physical AI refers to AI that combines physical laws and data-driven learning to understand and predict the real world, producing more accurate outcomes.

Unlike generative AI, Physical AI can be implemented in reality through humanoid robots, factories, and other industrial devices. Physical AI is praised for minimizing the gap between the virtual and real worlds to enhance result accuracy. Currently mainstream generative AI answers user queries to create creative content and provide services in virtual spaces. In this way, Physical AI is a more advanced AI concept than generative AI.

Humanoid robot unveiled at CES 2025
Photo by Yonhap News

American AI semiconductor company NVIDIA has introduced Physical AI as the next step following generative AI. Jensen Huang, NVIDIA’s CEO, defined the ultimate future of AI as Physical AI during his keynote speech on December 6 (local time), the day before the opening of the world’s largest consumer electronics and information technology exhibition, CES 2025, and unveiled a new AI model called "Cosmos."

CEO Huang described Cosmos as "a platform that generates a large amount of physically plausible synthetic data for training robot AI models." He added, "Using Cosmos significantly reduces both the time and capital required for robot learning."

Cosmos is a development platform composed of the World Foundation Model (WFM), tokenizer, guardrails, and video processing pipeline. While NVIDIA’s digital twin platform "Omniverse," released in 2021, focused on virtualizing static structures such as buildings, Cosmos simulates and generates real-world movements and motions. It can process 20 million hours of video in just 14 days, drastically reducing the work period that would take over three years using only a central processing unit (CPU). The tokenizer is known to be 12 times faster than previous versions. The tokenizer is a tool that converts images and videos into tokens (the smallest units of a corpus).

Jensen Huang, CEO of NVIDIA, delivering the keynote speech at CES 2025
Photo by Yonhap News

Humanoid robots are considered representative models of robots. In his keynote, CEO Huang introduced 14 companies, including Figure AI, Neura Robotics, and Hillbot, which will showcase humanoid robots alongside the unveiling of Cosmos. He emphasized, "The ChatGPT moment for robots is approaching," and stated, "Cosmos will be fundamental in advancing the development of robots and autonomous vehicles."

As AI advances, the introduction of humanoid robots into various environments is expected to increase demand significantly. The American global investment bank Goldman Sachs forecasts that the global humanoid robot market will reach $38 billion (55 trillion won) by 2035, with shipments totaling 1.4 million units. Compared to estimates made about two years ago, the market size has increased sixfold, and shipments have quadrupled. The International Federation of Robotics also predicted that the global humanoid robot market will grow at an average annual rate of 71% between 2021 and 2030.

This content was produced with the assistance of AI translation services.